A 'Complete Reset' in Computer Graphics
NVIDIA DLSS 5 represents Jensen Huang's argument that real-time graphics is entering a new era built around generative AI. At SIGGRAPH 2026, Huang said computer graphics is undergoing a fundamental reset as games, film, robotics and factory digital twins all demand virtual worlds that behave more like the physical world. In his framing, achieving that level of realism increasingly requires AI to participate directly in how pixels and scenes are generated.
The through-line of the keynote was convergence. Huang traced a lineage from the 2001 programmable GPU, through CUDA in 2006, to the 2012 moment when CUDA GPUs 'made first contact' with AI, and to RTX in 2018, when AI came back to graphics through real-time ray tracing and DLSS. With DLSS 1, one out of every four pixels was inferred; today, he noted, entire patches of the screen are.
Graphics, simulation, and AI are no longer separate disciplines sharing a chip, they are one pipeline.
What DLSS 5 Changes for Developers
- Grounding beats promptingDLSS 5 uses the engine's own rendered frame (universally available and pixel-aligned) as the signal constraining what the generative model may change.
- Intent is enforcedInternal renderer buffers like albedo, surface normals, and lighting information train the model to enrich realism without altering the story the scene tells.
- Control is layeredDevelopers pick models per scene, set global structure and tone intensity, and mask individual props with independent sliders via engine-side controls.
- It extends, not replacesDLSS 5 adds a generation stage on top of existing pipelines, and a better render still gives it better grounding to start from.
DLSS 5: The Render Frame Is the Prompt
The centerpiece was DLSS 5, unveiled by Edward Liu, the technical leader behind DLSS. Liu identified the gap plainly: traditional rendering builds images from first principles (authored scenes, physics-based light, camera control to the millimeter) while generative models learned photorealism from the real world but remain slow, giant, and hard to control.
DLSS 5's answer is a division of labor: the renderer keeps building the world exactly as the game authored it, and a learned generation stage enriches its appearance in real time.
The key insight is what grounds the model. Rather than steering generation with text prompts or reference images (signals that always leave the model room to improvise) DLSS 5 uses the frame the game just rendered. That signal is universally available across every engine, pixel-aligned with the shot, and dense enough to anchor character identity, lighting, mood, and composition.
Getting there required solving three hard problems: preserving artistic intent (a hero's scar may gain better skin shading, but it must never disappear), temporal stability frame by frame using motion vectors, and speed, because a 4K frame has roughly 8 million pixels and less than 16 milliseconds of budget at 60 fps.
Speed came from distillation. Nvidia compressed what giant foundational models know about world appearance into a small one-step pixel-space diffusion transformer that does exactly one thing: make real-time rendering look more real. It retains broad world knowledge (it understands characters, lighting, and context) but gives up generality; it cannot make funny cat videos anymore.
Nvidia says it is shipping this fall, and Liu was careful about the limits: the team is not claiming real-time photorealism is solved, and animation is not yet uplifted to the same degree as appearance.
Artists Direct, the Model Obeys
What distinguished the demo was who ran it. Nvidia brought a creative artist on stage to direct DLSS 5 live, making the argument that the mapping from a rendered frame to a more realistic version is not one-to-one (multiple outcomes could be correct) so decisions about which look to apply, where, and how strongly belong to the people making the game, not the model. 'Artistic direction now becomes the new ground truth, ' Liu said.
The controls shown ranged from broad to granular. Developers can choose among different trained models per scene or cutscene, set global structure intensity (high-frequency detail like ambient occlusion and subsurface scattering) and tone intensity (expressive lighting), and use the model's own semantic understanding to apply enhancement only to the base character. Engine-side masks then let them target individual props (a pitcher, grapes, bottles) each with independent sliders.
In one foliage scene, turning DLSS 5 on added subsurface scattering to plants, an effect traditionally so expensive it would be impractical in real-time rendering.
The positioning matters for how the industry reads this moment. Liu placed DLSS 5 as a third category of AI in real-time rendering, beyond reconstruction (DLSS supersampling) and function approximation (neural caches and materials): generation, where the model actually changes the final image.
He also added a second scaling lever to the historical curve, image quality can now improve not just by scaling the render but by scaling the model or improving training, at no additional runtime cost. The framing was consistent throughout: not AI replacing graphics, but AI extending graphics, the way programmable shaders and ray tracing did before.
AI Physics: A Million-to-One Compression of Reality
The keynote's second act moved from images to physics. Neil Ashton showcased one of the largest and most accurate Earth-climate simulations ever achieved (a coupled atmosphere, land, and ocean model running on more than 20, 000 GPUs at one-kilometer resolution, a 50-billion-cell grid that won the Gordon Bell Prize. But such fidelity is computationally punishing, and running at that scale every day is not feasible.
The alternative Nvidia has helped pioneer: train AI models on simulation data, then predict autoregressively in seconds or minutes what once took hours or days) at accuracy equal to or better than traditional simulation, now in production at weather centers worldwide.
The engineering numbers were the most striking claim of the session. An open dataset of roughly 2, 000 aircraft simulations at different geometries and boundary conditions, run with more than a billion degrees of freedom, generates around 200 terabytes of data.
Tokenized through a transformer architecture, that compresses to a model checkpoint of only about 200 megabytes (a million-fold reduction) and the trained model predicts flow for unseen geometries more than 10, 000 times faster than simulation, within one or two percent accuracy, on a single modest RTX 6000 GPU.
Ashton's closing pitch was aimed squarely at the graphics community in the room: the differentiable physics and geometry-modeling techniques that made this possible largely came out of SIGGRAPH, and Nvidia wants that expertise applied to electromagnetics, structural simulation, and beyond, potentially discovering designs that could never be achieved before.
Cosmos: When Compute Becomes Data
The final act connected everything to physical AI. Minguk Kang, VP of Nvidia's Cosmos Lab, opened with the field's core bottleneck: 'Physical AI has a data problem.' Robots need real-world experience, but the real world only runs in real time, you cannot fast-forward reality. Cosmos, a world foundation model, reframes the constraint: if a model can generate worlds from a robot's perspective, data stops being something you collect and becomes something you compute.
Compute is data.
Cosmos 3 unifies world generation, understanding, and action in one omnimodel built on a mixture-of-transformers architecture: an autoregressive tower that reasons and a diffusion tower that generates, aligned through shared multimodal attention. A common geometric action vocabulary (ego pose, hand vectors, graspers) lets one model span cars, arms, humanoids, and even human egocentric video, turning ordinary first-person footage into robot training data.
The new Cosmos 3 is a 4-billion-parameter model built for one purpose: running real time on devices, so robot policy and video analytics happen where the robot lives with no round trip to a data center.
The demo that tied it together was Cosmos Dreams, a neural closed-loop simulator for autonomous vehicles: a policy model takes an action, and Cosmos generates photorealistic, frame-by-frame sensor output conditioned on that action. The hardware story was the headline (when Nvidia first built the real-time demo it needed 64 B300 GPUs, then brought it to 16 Vera Rubin GPUs, and the new Cosmos model now runs it on a single RTX Pro 6000 workstation GPU.
For AV developers, that turns a datacenter-scale effort into something drivable from a PlayStation controller on a desk) with Kang noting Cosmos models rank number one across eight benchmarks spanning generation, reasoning, and policy, and Cosmos 3.5 already in development.